dqn_heads = []
act_index = np.zeros(num_heads, dtype=int)
epsilon_eval = 0.05
done = False
for i in range(num_heads):
the_model = torch.load('./tmp/model_epoch_200/model' + str(i) + '.pth')
dqn_heads.append(DQN(num_actions))
dqn_heads[i].load_state_dict(the_model)
#dqn = DQN()
#dqn.load_state_dict(the_model)
with torch.no_grad():
var_phi = autograd.Variable(torch.Tensor(1, 4, 84, 84))
while (not done):
for i in range(num_heads):
act_index[i] = sample_action(atari,
dqn_heads[i],
var_phi,
epsilon=epsilon_eval,
num_actions)
act_index_vote = np.bincount(act_index).argmax()
phi_next, r, done = atari.step(VALID_ACTION[act_index_vote])
atari.display()
time.sleep(0.01)
from model import DQN
from pong import Pong
import torch
import torch.autograd as autograd
from utils import sample_action
the_model = torch.load('./tmp/model.pth')
dqn = DQN()
dqn.load_state_dict(the_model)
pong = Pong()
done = False
VALID_ACTION = [0, 2, 5]
var_phi = autograd.Variable(torch.Tensor(1, 4, 84, 84), volatile=True)
while(not done):
act_index = sample_action(pong, dqn, var_phi)
phi_next, r, done = pong.step(VALID_ACTION[act_index])
pong.display()
